Drowsiness Detection System

A real-time driver drowsiness detection system that monitors eye state through a webcam and triggers an audio alarm when prolonged eye closure is detected.

How It Works

1. Captures a live webcam feed using OpenCV
2. Detects the face and eyes using Haar Cascade classifiers
3. Classifies each eye as Open or Closed using a pre-trained CNN
4. Maintains a running score — +1 each frame both eyes are closed, −1 otherwise (floor 0)
5. When score exceeds 15, the system:
   • Plays alarm.wav via pygame
   • Draws a pulsing red border on the video feed
   • Saves a snapshot of the drowsy frame as image.jpg

Press q to quit the detection window.

Project Structure

drowsiness-detection/
├── drowsiness detection.py       # Real-time detection pipeline
├── model.py                      # CNN training script
├── alarm.wav                     # Alert sound
├── image.jpg                     # Snapshot saved on drowsiness event
├── models/
│   └── cnnCat2.h5                # Pre-trained CNN weights
└── haar cascade files/
    ├── haarcascade_frontalface_alt.xml
    ├── haarcascade_lefteye_2splits.xml
    └── haarcascade_righteye_2splits.xml

Requirements

• Python 3.12+
• uv (recommended package manager)
• A webcam

Installation

# Clone the repository
git clone <repo-url>
cd drowsiness-detection

# Install dependencies
uv sync

Usage

Run real-time detection

uv run python "drowsiness detection.py"

The detection window will open. Press q to quit.

Note: The script uses cv2.VideoCapture(1) by default. If your webcam is not detected, change the index to 0 in drowsiness detection.py line 16.

Retrain the model (optional)

The pre-trained model (models/cnnCat2.h5) is included — retraining is only needed if you want to use a custom dataset.

Prepare your dataset with this structure:

data/
├── train/
│   ├── Open/
│   └── Closed/
└── valid/
    ├── Open/
    └── Closed/

Then run:

uv run python model.py

The retrained model will be saved to models/cnnCat2.h5.

Compatibility note: drowsiness detection.py uses tf_keras (TF 2.x standalone package), while model.py uses the older keras API. If you retrain, update model.py imports to tf_keras to avoid conflicts.

Model Architecture

Input: 24×24 grayscale eye region of interest (ROI)

Conv2D(32) + MaxPool → Conv2D(32) + MaxPool → Conv2D(64) + MaxPool
→ Dropout(0.25) → Flatten → Dense(128) → Dropout(0.5) → Dense(2, softmax)

Output: 0 = Closed, 1 = Open

Dependencies

Package	Purpose
opencv-python	Webcam capture, Haar Cascade detection, frame rendering
tf-keras	Loading and running the pre-trained CNN
tensorflow	Backend for tf-keras
numpy	Image array manipulation
pygame	Audio alarm playback